This is a—very rough—translation of Max Bense’s essay “On Natural and Artificial Poetry” (“Über natürliche und künstliche Poesie”) from his 1962 collection Theory of Texts (Theorie der Texte, the plural is intended). It is not only prescient for its distinction between “human” and “machine-generated” text, but already tackles the problem of intention we are confronted with in the context of large language models today. I employ his distinction in my essay “On Artificial and Post-Artificial Texts.”
(I do not hold the rights to the original piece; copyright holders, please contact me should I be infringing your rights.)
(Image: Max Bense in the TV studio of Bayerischer Rundfunk 1966. Photograph: P. Sessne, from the University of Stuttgart Website.)
To elucidate a general concept of poetry, it may be helpful first to make a distinction between natural and artificial poetry [natürliche und künstliche Poesie]. In both cases one works with words—with their derivatives, which can be interpreted as deformations with respect to an underlying word space, and their sequences, which are arranged linearly or as a plane. For the aspects we are concerned with here, however, the difference in the way they are brought about [Art der Entstehung] is essential.
Natural poetry is understood here as the kind of poetry—this is the classical and traditional case—that has as its prerequisite a personal poetic consciousness, as Hegel already called it; a consciousness that possesses experiences, feelings, memories, thoughts, conceptions of an imagination, etc., in short, a pre-existent world, and is able to give it linguistic expression. Only within this ontological framework can there be a lyrical I [lyrisches Ich] or a fictive epic world. The poetic consciousness in this sense is principally concerned with transpositions—it puts being into signs. The epitome of these signs we call language insofar as they, meta-linguistically, have an ego relation and a world aspect. In natural poetry, writing does not cease to be an ontological continuation: Every word that it expresses follows the world experiences of an I, and even the aesthetic rank that is given to it could still be understood as a reflex of this world.
Artificial poetry, on the other hand, is understood here as a kind of poetry in which—insofar as it was produced, for instance, by a machine—there is no personal poetic consciousness with its own experiences, feelings, memories, thoughts, imaginations, etc., that is, no pre-existent world, and in which writing is no longer an ontological continuation through which the world aspect of the words could be related to a self. As a result, neither a lyrical I nor a fictive epic world can be meaningfully extracted from the linguistic fixity of this poetry. Thus, while for natural poetry an intentional beginning of the semantic process is characteristic, for artificial poetry, there can only be a material origin.
Of course, the differences mentioned are primarily ideal types. Only approximations, one must assume, really exist. For example, due to the precision with which rhythm and meter are handled, material features of artificial poetry can also appear in intentional natural poetry. As far as the examples of actually existing artificial poetry are concerned—which, for instance by machine selection, let the word sequence proceed in an exclusively material and successive manner—it is, as I said before, advisable to speak simply of “texts,” in order to indicate with this term the generalized form of poetry, which is attainable in them. “Text,” in this context, denotes any sequence or arrangement of words that emerges selectively and contingently from and in relation to an underlying textual space (vocabulary), and that allows for certain deformations of the individual word.
There are mainly three types of programs for the realization of such texts in artificial poetry: statistical, structural, and topological ones. The program is statistical if it uses certain frequency distributions of words to form the selected word sequence, that is, if it programs these frequency distributions; the program is structural if it structures the selected word sequence (macro-linguistically), that is, if only certain word classes, verbs, nouns, adjectives etc. are allowed in it or if certain arrangements of the selected words on the surface are predetermined; and we call the program topological if words are selected on the basis of neighborhood relations or deformations (or deformation classes) that are previously determined.1
The programs thus start from frequency properties, from class and arrangement properties, and from neighborhood and deformation properties. Characteristic texts of statistical programming are for instance Shannon’s well-known text approximations. The selection of the word sequences into text segments proceeds stepwise, by falling back in each case on repertoires of different frequency distributions, which correspond however more and more to a real text. The text consists thus of individual text segments, the first of which appears completely arbitrary and the last of which appears least arbitrary. Thus, we are dealing with stochastic text sequences that exhibit the statistical structure of Markov chains. Characteristic texts of a structural program are the serial word sequences of Gertrude Stein and Helmut Heissenbüttel or the flat word constellations of Eugen Gomringer’s Concrete Poetry and the Brazilian “noigandres” group. Finally, characteristic texts of a topological program can be found, as far as the selection of certain neighborhood relations is concerned, in the di-, tri-, and tetragram selections of Shannon’s approximations as well as in that of serial (Gertrude Stein) and concrete poetry (“noigandres”). Texts whose aesthetic programming is consciously based on text-topological deformation have been published by the author. Of course, these types of programming can only be kept distinct as ideal types. In the context of the actual realization of artificial poetry—whether by human or machine writing (in the process of selection)—one programming type may be predominant, but it immediately involves at least one additional one.
We can call such programs material programs, insofar as their “themes” belong entirely to the intrinsic world of the material, that is, to the text space consisting of words as elements. This is clearly true of the formations of “sequences,” “series,” “approximations,” “frequencies,” “classes,” “deformations,” etc. The resulting poetry can also be called mathematical, as far as the programming proceeds mathematically, that is, is governed by mathematical determinants like “sequences,” “classes,” “frequencies,” “sets,” “subsets,” “Boolean lattices” etc. The term “cybernetic” is appropriate for artificial poetry when the selection—the “writing”— has been done by machine, that is, with the aid of program-controlled electronic computing devices. Examples of specifically cybernetic artificial poetry have been published in issue 6 of the journal “rot.”
It now seems possible to me to point out further differences between natural and artificial poetry. In natural poetry, certain classes of words—nouns, verbs, and adjectives, for instance—enjoy a certain privileged position with respect to the formation of the semantic content, which usually appears as the carrier of the aesthetic. In artificial poetry, on the other hand, in which the material realization of the words or the sequence of words coincides with the aesthetic one and the semantic content is disregarded, all words are a priori equal. Furthermore, the following is to be stated: Natural poetry can and must be interpreted, because it is mostly only with the interpretation that the ego relation of the words on the one hand and their world aspect on the other hand can be apperceived and the communicative process can only come to a conclusion in this way. Since the essence of interpretation consists primarily in the production of the ego relation and the world aspect of a text, in the recourse to what we have called “ontological continuation,” and since such continuation does not exist in artificial poetry, interpretation has no meaning for it. Artificial poetry thus contains in general much more non-communicative components than natural poetry, at least as far as the realized information is concerned.
We thus come up against the aesthetic discussion of our problem. Obviously, every aesthetic realization is characterized by specific non-communicative components. Most likely, they are precisely characteristic for the construction of the original aesthetic information. In information aesthetics, it is shown that aesthetic information—unlike semantic information—is not encodable, only realizable. One can say that in artificial poetry, pieces of aesthetic information appear as pure realization, since in the ideal case they are not given semantic carriers in the usual sense (statements, ideas, etc.). In any case, artificial poetry is possible as pure, absolute poetry, insofar as one cannot assume it to have any prefixed meaning of an evocative character. Artificial poetry has, like numbers, only an existential, not an essential, power; it realizes the words and their connectives as linguistic materials, not as linguistic carriers of meaning. With this, however, it is absolutely clear that artificial poetry is, in principle, pure realization poetry in consequence of its non-communicative components. For only meaning is transmittable (and therefore encodable), but not realization.
Thus, an interpretation can never be more than the derivation of a piece of information from a piece of information, namely from its communicatively accessible, semantic components. Interpretation refers to redundant features. The purpose of interpretation is to trace back a selectively realized piece of information to its redundant (high-frequency) building blocks, to its carriers of meaning. This process of tracing-back to meaning is the reductive function of every interpretation. An interpretation of the aesthetic components of a piece of information (an aesthetic interpretation) would have to be the production of semantic information from aesthetic information, thus the transfer of innovative (low-frequency) features of the selectively realized information into redundant (high-frequency) ones, which would be equal to a cancellation of the aesthetic information. Thus, a piece of aesthetic information could be interpreted, according to its being [seinsgerecht], only by deriving another piece of aesthetic information from it. We can only gain the concept of aesthetic interpretation in the actual sense here, however, and cannot investigate it any further. There is only one more thing I would like to point out. Of course, there are word sequences in the artificial poetry (for example, in the Markoff chain texts) that allow for meaning. Werner Meyer-Eppler spoke of “real text” [wirklicher Text] in this context. Artificial poetry can certainly take on the features of natural poetry.
For the concept of interpretation, this results in the distinction between voluntary and involuntary meaning [willkürliche und unwillkürliche Bedeutung]. The former depends on an interpreting will, the latter does not; the former was predetermined [präfixiert], the latter was not. In this sense, every fictional epic world is also voluntary like the real world of the lyrical subject. (Cf. Käte Hamburger, Logik der Dichtung, 1958.) But whatever in artificial poetry may indicate an epic-fictional or lyric-real world, both remain involuntary. The mode of involuntariness distinguishes the world that can be prepared from artificial poetry from the world that can be prepared from natural poetry.
Translation Hannes Bajohr
Source: Max Bense, “Über natürliche und künstliche Poesie,” in: Theorie der Texte (Cologne: Kiepenheuer & Witsch, 1962), 143–147.